Electromagnetic Inverse Scattering from a Single Transmitter (CVPR 2026 Highlight)

Yizhe Cheng, Chunxun Tian, Haoru Wang, Wentao Zhu, Xiaoxuan Ma, Yizhou Wang

🌐 Project Page | 📄 Paper | ▶️ Video

This repository contains the PyTorch implementation of Electromagnetic Inverse Scattering from a Single Transmitter, with unified training and evaluation for MNIST, Circular-cylinder (CYLINDER), Institut Fresnel (IF), 3D MNIST, and 3D ShapeNet.

TODO

• Release unified training and evaluation code.
• Provide configurations for the experiments presented in the paper.
• Release the training datasets.
• Release data generation and preprocessing code.

1. Introduction

We propose a fully end-to-end, data-driven framework for electromagnetic inverse scattering. By learning data distribution priors, the model compensates for information scarcity in sparse-transmitter setups and predicts relative permittivity directly from scattered-field measurements and spatial coordinates.

The shared MLP takes the real and imaginary measurements, together with Fourier-encoded query coordinates, and predicts the permittivity at each pixel or voxel.

2. Preparation

Environment setup

Python 3.10 or newer is required. For example:

conda create -n GenEISP python=3.11.10
conda activate GenEISP
pip install -r requirements.txt

Dataset Setup

1. Download the test datasets
   Please download the test datasets from GoogleDrive.

2. Extract and organize the data
   Place test.zip and all ten training parts (train.zip.001–train.zip.010) in ../downloads/, then run:

   python prepare_data.py --download-dir ../downloads --output ../data

   The folder structure should look like this:

   data/
   ├── train/
   │   ├── cylinder_mnist_inc16/
   │   ├── if_inc8/
   │   └── if_inc18/
   └── test/
       ├── cylinder/cylinder_test_inc16/
       ├── mnist/mnist_test_inc16/
       └── IF/
           ├── FDE/
           ├── FDI/
           └── FTD/
   

   Update the dataset paths in the configuration file if your filenames differ.

3. Training data
   Download all training parts from GoogleDrive. MNIST and CYLINDER use a combined training set; IF uses matching synthetic training data.

Pre-trained Model Setup

1. Download the pre-trained model
   Download the available model packages from GoogleDrive.

2. Organize the model weights
   Extract the model packages into ../ so checkpoints are placed in ../runs/:

   python -m zipfile -e ../downloads/models_2d.zip ..
   python -m zipfile -e ../downloads/models_3d.zip ..

   These checkpoints are for evaluation. For training initialization, use --weights, not --resume.

3. Evaluation

⚙️ 3.1. Configuration File (config.yaml)

Configuration files follow the naming convention:
[dataset_name]_noise[level]_N[transmitter_number].yaml.

Example File	Meaning
cylinder_noise05_N16.yaml	Cylinder, 5% noise, 16 transmitters
mnist_noise30_N16.yaml	MNIST, 30% noise, 16 transmitters
mnist_noise05_N1.yaml	MNIST, 5% noise, 1 transmitter
IF_FDE_noise00_N8.yaml	FoamDielExt, no added noise, 8 transmitters
IF_FDI_noise00_N8.yaml	FoamDielInt, no added noise, 8 transmitters
IF_FTD_noise00_N18.yaml	FoamTwinDiel, no added noise, 18 transmitters
3Dmnist_noise05_N6.yaml	3D MNIST, 5% noise, 6 transmitters
3DShapeNet_noise05_N1.yaml	3D ShapeNet, 5% noise, 1 transmitter
mnist_noise15_N16.yaml	Noise-level ablation, 15% noise
mnist_noise30_N16_data25.yaml	Training-data ablation, 25% training data

Each configuration contains train and test sections. Set train.channel: 0 for one transmitter or train.channel: -1 for all transmitters.

▶️ 3.2. Run Evaluation

• For 2D datasets (cylinder, mnist, IF):

python test.py --config config/cylinder_noise30_N16.yaml
python test.py --config config/mnist_noise30_N16.yaml
python test.py --config config/IF_FDE_noise00_N8.yaml

• For 3D datasets (3D MNIST, 3D ShapeNet):

python test.py --config config/3Dmnist_noise05_N1.yaml
python test.py --config config/3DShapeNet_noise05_N1.yaml

4. Training

Update train.train_data, train.test_data, and train.output in the selected configuration, then run:

python train.py --config config/mnist_noise30_N16.yaml
python train.py --config config/IF_FDE_noise00_N8.yaml
python train.py --config config/3Dmnist_noise05_N1.yaml

MNIST and CYLINDER configurations with the same noise level and transmitter count share one model. Train it once, then evaluate both datasets.

• Noise-level ablation:

python train.py --config config/mnist_noise15_N16.yaml
python test.py --config config/mnist_noise15_N16.yaml

• Training-data ablation:

python train.py --config config/mnist_noise30_N16_data25.yaml
python test.py --config config/mnist_noise30_N16_data25.yaml

train_fraction reduces training data only; evaluation uses the complete test set.

Citing

If you find this work useful for your research, please consider citing:

@inproceedings{cheng2026electromagneticinversescatteringsingle,
  author    = {Yizhe Cheng and Chunxun Tian and Haoru Wang and Wentao Zhu and Xiaoxuan Ma and Yizhou Wang},
  title     = {Electromagnetic Inverse Scattering from a Single Transmitter},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}